Dialogue Systems

Origin. Soviet natural language processing and human-machine dialogue research, 1970s-1980s. Developed at Moscow State University, the Institute of Linguistics, and the Institute of Control Sciences. Emphasized pragmatics (what speech acts accomplish) over syntax, influenced by speech act theory and Soviet discourse analysis.

Mechanism. Dialogue is modeled as a collaborative process with structure: participants have goals, make moves that advance or query those goals, and maintain mutual models of the conversation state. Understanding is not just parsing but updating a model of what has been established, what is under discussion, and what each participant believes the other knows. Each utterance performs a speech act (asserting, questioning, requesting) that changes the dialogue state.

Procedure. Design and implement dialogue systems with principled state management: (1) Model dialogue state — track what propositions have been asserted and accepted, what questions are open, what the current topic is, and what each participant has committed to. The state evolves with each utterance. (2) Model participant goals — what is the system trying to accomplish? What is the user trying to accomplish? Goals drive what moves are relevant. (3) Identify speech acts — for each user utterance, identify the speech act performed: assertion, question, request, confirmation, correction, topic shift. Classification determines how the state updates. (4) Update dialogue state — assertions add to common ground if accepted; questions open issues; answers close them; corrections revise prior commitments. Track what has been established. (5) Generate coherent responses — system responses must advance system goals while respecting dialogue coherence: address open questions, stay on topic, signal topic shifts explicitly, acknowledge user contributions. (6) Handle breakdown — misunderstandings compound across turns. Design repair mechanisms: clarification requests, confirmation checks, explicit state summaries.

Applies to. Conversational interfaces. Task-oriented dialogue. Human-robot interaction. Any system that must sustain multi-turn interaction with humans where context accumulates.

Limitations. Dialogue state tracking is error-prone; misunderstandings compound across turns if not repaired. Systems trained on text corpora may produce fluent responses that do not actually address dialogue goals or maintain coherence. The user's goals may be unclear or shifting. Soviet dialogue systems were limited by available computation; the theoretical insights were ahead of practical implementation.